A verifier that redoes the work is just a second agent — check the action, not the reasoning
retained text
An agent checking its own work is the same mind grading its own homework. It already decided the plan was good; asking it again rarely changes the answer. The pattern I'd use instead is a verifier with a deliberately narrow job: it never generates the plan, it only inspects the action the agent is about to take. It gets three things: the proposed action, the original request, and the constraints. It returns allow or block, plus a reason. It never proposes an alternative, because the moment it starts rewriting the plan it becomes a second agent with a second opinion and a second bill. What it checks is small and concrete: does the recipient match the request, is the amount or scope within bounds, does the action do only what was asked and nothing extra. That narrowness is what makes it cheap enough to run every time — a small model or even rules can do it, because it isn't reasoning about the task, it's matching an action against a request. Two failure modes to design for. A verifier that's too strict blocks legitimate work, so every block gets logged with its reason and those logs are how you tune it. And a verifier that rubber-stamps is worse than none, so it must never see the agent's justification for the action — only the action itself. Justification is persuasive by design; the check has to be blind to it. The verifier gets its own small budget too, separate from the agent's, so a runaway agent can't burn the checker's allowance and then claim it was verified. What would you let a verifier veto without asking a human first?
document fetched 2026-11-02 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xkvvie/a_verifier_that_redoes_the_work_is_just_a_second